Skip to content

feat: model baseline context evaluation#64

Merged
N4M3Z merged 1 commit into
mainfrom
feat/fable-context-eval
Jun 16, 2026
Merged

feat: model baseline context evaluation#64
N4M3Z merged 1 commit into
mainfrom
feat/fable-context-eval

Conversation

@N4M3Z

@N4M3Z N4M3Z commented Jun 16, 2026

Copy link
Copy Markdown
Owner

Summary

An OptimizeContext run against the Fable 5 baseline: drop or offload rules the model now follows without being told, and move always-on reference into the skills that own it.

  • Rules: offload DeadVariables, NeverRefuseWork, PrepareDataOutsideLoops to rules/claude/claude-opus-4-7/ + claude-opus-4-8/ (parked until forge-cli#60 lands model-level qualifiers); slim BehavioralSteering to the rows the model can't infer; delete ScanRepository (VerifyClaims covers it). Stale-fixes to UseRTK, UseForge, KnownIssues, RuleStructure, NoTabs, MakefileFirst, SelfLearning, GitWorktrees, KeepFormatting, LineEndings, MemoryFiles.
  • Rule into skill relocations (drop the always-on rule, fold it into the skill that already owns the topic): AgentStructure → BuildAgent, ArtifactComposition → BuildModule, ClaudePluginStructure + ClaudeCoworkLimitations → BuildPlugin, DeployManifest + DriftComparison + ProvenanceVerification → ProvenanceAudit, SecretScan + GitCryptUserData → SecretScan, WindowsConstraints + HookScripts + HookQuality → BuildHook, ValidationChain → VersionControl.
  • CrossProviderAssemblyNoVendorLockIn: reframed around the point (no AI vendor is a single point of failure), "harness" not "provider".
  • New: Deslop rule, OptimizeContext skill, config/models.yaml, MVPR-0004 writing up the method. MVPR-0003 + PROV-0005 amended.
  • Skill body sweep: slims and stale-fixes across BashConventions, MarkdownConventions, BuildModule, BuildAgent, ArchitectureDecision, and more. BuildHook is past slimming and needs a rewrite, filed as forge-core#59.
  • Qualifier dirs to exact model IDs: claude/opus-4-6claude/claude-opus-4-6.

Other sessions' work braided into the same jj change is held back for its own PRs: the jj/Entire tooling, the OnePurpose / UserConfig / VerifyDecisions / DataWorkingLayer / ScratchDirectory rules, a few unrelated skill adds.

Test plan

  • forge validate --source . green at each batch through the run
  • CI on the isolated branch
  • post-merge: make install + /reload-skills

@N4M3Z N4M3Z requested a review from LangrMartin as a code owner June 16, 2026 23:20
@N4M3Z N4M3Z force-pushed the feat/fable-context-eval branch from 06cf29a to f3ee39b Compare June 16, 2026 23:43
@N4M3Z N4M3Z merged commit ed2192f into main Jun 16, 2026
1 check passed
@N4M3Z N4M3Z deleted the feat/fable-context-eval branch June 16, 2026 23:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant